JOURNAL OF COMMUNICATIONS, VOL. 4, NO. 7, AUGUST 2009
487
Automatic Animal Tracking Using Matched Filters and Time Difference of Arrival Robert MacCurdy Cornell University/Mechanical and Aerospace Engineering, Ithaca, USA Email:
[email protected]
Rich Gabrielson, Eric Spaulding, Alejandro Purgue, Kathryn Cortopassi, Cornell Laboratory of Ornithology/ Bioacoustics Research Program, Ithaca, USA
Kurt Fristrup National Parks Service, Fort Collins, USA Abstract—A method for tracking animals using a terrestrial system similar to GPS is presented. This system enables simultaneous tracking of thousands of animals with transmitters that are lighter, longer lasting, more accurate and cheaper than other automatic positioning tags. The technical details of this system are discussed and the results of a prototype are shown. IndexTerms—radio tracking, TDOA, wildlife tracking, CDMA, localization, pseudolite, GPS, matched filter I. INTRODUCTION Radio-tracking has been widely used to monitor wildlife movements since the 1960s [1,2] with hundreds of scientific papers published using some variant of this method. For the majority of these studies, an operator monitors received signal strength while changing the orientation of a directional receiving antenna. The direction yielding the maximal signal strength is recorded as a pointing vector to the tagged animal. This simple method is adequate to guide a researcher to the location of a focal individual, and triangulation using two or more receiving stations can be used to track a few individuals simultaneously. The accuracy of this method is detailed in [3]. However, this method yields relatively few position fixes per hour, and fully absorbs an operator's attention and effort. Automatic tracking systems have been developed using fixed receiving towers [4], [5], [6]. Most efforts involve directional antennas and rely on the beam pattern of the antennas to infer a direction of arrival. These approaches generally suffer an error in the 1-10 degree range, depending on the implementation [6] and the cross-bearing positional error for each receiving station increases linearly with range. In addition to terrestrial radio-tracking methods, two satellite based options exist: GPS and Argos. These two systems provide location information using different techniques. GPS employs a network of orbiting satellites that broadcast signals to a terrestrial receiver which uses a Time Difference of Arrival (TDOA) algorithm to estimate its position. The Argos system exploits the frequency shift in the received transmission (Doppler shift) caused by a satellite’s motion relative to a Manuscript received December 14, 2008; revised May 12, 2009; accepted June 15, 2009.
© 2009 ACADEMY PUBLISHER
transmitter on the earth. The principle drawback of GPS is that it does not directly provide a means of reporting position information back to the researcher. The position information is either stored and retrieved later, or downloaded via an auxiliary radio frequency link. Weight is also a limiting factor for satellite-based systems. The typical maximum allowable tag to body weight ratio is 5%, with lower being preferable. The smallest GPS tags at present are in the 22 to 150 gram range, which limits their application to larger animals (>440 g). The Argos system can achieve reasonably good accuracy; the best service available advertises 150 m accuracy. However, this level of accuracy is not often available, and the other three accuracy classes range from 150 to 1000 m. Argos tags are generally smaller than GPS tags, but still only allow tracking of animals heavier than 200 g. This weight constraint excludes 40% of all bird species [7]. The cost of GPS and Argos tags is also prohibitive for large-scale studies. The complexity and low volume of these tags lead to typical single unit prices in the $1500 to $4000 range, with little cost reduction at larger volumes. The weight, cost, and performance of existing radiotracking techniques necessitate a new approach. We have designed, built and installed a prototype system based on Code Division Multiple Access (CDMA) and Time Difference of Arrival (TDOA) that is capable of automatically locating thousands of tags in real time. This system borrows concepts from the GPS system in general, and from pseudolites in particular. Pseudolites (or pseudo-satellites) refer to terrestrial devices which transmit GPS signals. An excellent summary can be found in [8]. Pseudolites have existed for many years in various forms; in fact some of the earliest pseudolite tests actually predate the launch of GPS satellites [9]. II. SYSTEM DESIGN A. System Overview Previous attempts at terrestrial automatic wildlife radio tracking have relied on conventional, fixed frequency, narrow band transmissions. These transmissions carry relatively little information for the power consumed during transmission. In contrast, our approach relies on a new, broadband transmitter that emits a transmission
488
JOURNAL OF COMMUNICATIONS, VOL. 4, NO. 7, AUGUST 2009
modulated with a unique pseudo-noise (PN) sequence. This approach is similar in many ways to a GPS system operating solely in acquisition mode, however in our system, the mobile device to be tracked is the transmitter, rather than the receiver. The transmitted sequence occupies a typical bandwidth of 2MHz and lasts about 3 msec. Each transmitter operates on the same carrier frequency, which is currently 140MHz, but each transmitter modulates that common carrier with a different PN code. We have chosen Gold Codes for the system, due to their attractive cross-correlation properties. This yields a code division, multiple access system: many transmitters can operate simultaneously without significant interference. The receiver, unlike a conventional animal tracking receiver, “listens” continuously to the 2MHz wide bandwidth that the transmitted signal occupies and runs a matched filter detector. When a signal is detected, the receiver precisely timestamps the event and sends a data packet to a server with the event ID, tag ID, time and other information. The server accumulates these individual data packets from different receivers, which are located in different positions, and groups receive events. When multiple receive events correspond to the same transmission event (the tag was “heard” at multiple receivers), the server computes a position estimate. Three or more receivers must receive the same transmission in order for the server to compute an accurate position fix. If only two receivers hear the transmission, the transmitter position can be limited to a particular hyperbola, and if only one receiver hears the transmission, a simple presence/absence data point is available. The system uses precise timing of receive events to calculate transmitter position. One major benefit of this approach is that the accuracy of the system does not degrade as a direct function of distance, unlike the Direction Of Arrival (DOA) based systems. Significant power savings, relative to conventional techniques, are also realized by this approach. Very short and infrequent transmissions are possible because the receivers are listening constantly, and can listen for transmissions from all tags simultaneously. Conventional radio tracking tags transmit for 30 msec and repeat every 2 to 3 seconds. This transmission length and interval are the minimum allowable times for a human operator to accurately discern a reception event and determine tag direction. However, in our system the tag can be configured to transmit for only 3 msec and repeat only as often as independent position fixes are required. For example, it is entirely possible to transmit once every 5 min, which reduces tag power consumption by a factor of 1500, relative to traditional tags. This dramatic reduction in consumption enables multi year studies with transmitters similar in mass to conventional tags. If maximum tag lifetime is not a priority, lower power consumption can also enable lighter tags. Our tags are easily programmed, in contrast to conventional radio tracking tags, which must be custom-built for a particular frequency and cannot be readily modified.
B. The Transmitter The transmitter (tag), shown in Figure 1, is based on an inexpensive, very low power microcontroller, along with a separate PLL, mixer and amplifier. Our design integrates off-the-shelf ICs in order to avoid the high cost and long development time of a custom ASIC. This choice results in an implementation that is larger than it could otherwise be, but this tradeoff allows rapid development. The tag’s 140MHz center frequency implies a ¼-wave antenna length of approximately ½ meter. This is too long for most small birds to manage, so the actual antenna used is often between 15 and 25 cm. Despite the efficiency penalty that these electrically short whip antennas impose; they are very common in animal tracking applications. The tag uses a binary phase shift keyed (BPSK) modulation scheme to directly modulate and spread the carrier power, as shown in Figure 2. The modulation rate is 1MHz, resulting in a 2MHz wide main band. The tag is programmable for center frequency, transmission interval, pseudo-noise code, chip-rate, RF output power, and operating schedule. This programmability allows tailoring the tag parameters to the application, which maximizes lifetime for a particular tag mass. The current tag mass is 1.4g without the battery and encapsulation. Future versions will reduce mass by 50%. Transmitting an RF signal is the most power consuming operation for any small transmitter. In conventional tracking systems, many of these transmissions are not used: receivers can only tune one tag at a time, human operators require multiple transmissions per bearing, and multiple bearings are required to determine a location. In contrast, the automatic detection algorithm running in our receivers is capable of determining tag location from a single transmission. This capability enables a very low duty cycle, dramatically increasing tag lifetime.
© 2009 ACADEMY PUBLISHER
Fig. 1 Transmitter board, with typical battery and programming tab still attached
Each tag sends two PN sequences, one after the other, per transmission interval. The first sequence is common to all the tags and serves as an acquisition and synchronization signal. The second sequence is unique to each tag and encodes ID. Each sequence is chosen from the family of Gold [10] codes available from an 11 bit
JOURNAL OF COMMUNICATIONS, VOL. 4, NO. 7, AUGUST 2009
489
generator. Gold codes are attractive due to their predictable cross-correlation performance and relatively easy generation. The choice to send two PN sequences was motivated by the limited processing power of the receiver stations. In principle, each tag could send only its unique ID sequence. This approach would require the receivers to search in code ID space as well as in code time-alignment. The DSP processing power available is insufficient for this task. Therefore, we employ a two stage detection process: search continuously and in real time for the acquisition and synchronization signal, and when it is detected, search in code ID space to identify the tag. This latter step need not take place in real time. This approach, and our current implementation, causes the system to block subsequent tag detections during the 3 milliseconds that the signal from the first detected tag occupies. This reduces the multiple access potential that true CDMA techniques offer. We reduce the likelihood that tags will repeatedly interfere with each other by introducing a random offset into their transmit intervals. If two tags interfere during one interval, they are very unlikely to interfere in the near future. Very short tag transmit durations also mitigate interference. A system employing all 2049 possible IDs, with each tag transmitting every minute, would only be using 10% of the available time. Systems with fewer tags or longer transmit intervals would experience reduced interference. The family of Gold codes we use supports 211+1 individual IDs operating simultaneously in a particular study area.
least several kilohertz of spacing, and the scan speed must be increased as the scanned bandwidth increases in order to avoid missing a transmission event. The latter requirement places a hard lower limit on the duration of the transmitted signal from a traditional tag, though future scanning receivers could address this problem by abandoning the traditional super-het topology in favor of a wideband, software defined architecture. These considerations indicate that both RF spectrum allocation and engineering effort will be more efficiently translated into tracking capacity using the CDMA approach. C. The Receiver Figure 3 shows a single receiver system, including the equipment box, tripod, batteries and antenna mast, which extends 5 meters above what is shown in the image. The receiver uses a 2-7/8λ phased element monopole antenna that yields approximately 6dB of gain. The system requires approximately 10 Watts and at 20 kg (not including lead acid batteries) can be easily moved. The receivers use a method similar to GPS in order to automatically locate animals wearing our tags. Each receiver runs a real time, matched filter detector that continuously seeks matches for tag PN codes. The receivers detect the initial synchronization code by performing an operation similar to that of a GPS receiver in acquisition mode. The GPS acquisition process
Fig. 2 Transmitter output showing unmodulated and modulated carrier
The system uses 2049 unique IDs, but identical tag IDs could potentially be assigned to more than one individual if the expected individual movement patterns and continuity of tracking are likely to provide decisive identification. For example, the same ID could be given to a tag on a rabbit and a tag on a wolverine, and there would be very little chance of confusion. A similar argument could be applied to animals whose usage of an area is clearly demarcated by season. The population of tags tracked by this system can be orders of magnitude larger than systems that utilize carrier frequency to identify individuals. Automatic tracking with narrow band systems does not scale well for two reasons: the available bandwidth will be rapidly consumed by tags whose crystal tolerances require at © 2009 ACADEMY PUBLISHER
Fig 3. Typical receiver setup
performs an ambiguity function search, which is a two dimensional search in offset frequency and code phase
490
JOURNAL OF COMMUNICATIONS, VOL. 4, NO. 7, AUGUST 2009
space. The animal tracking system simplifies the problem however by eliminating the requirement to search in frequency (Doppler) space, since the expected tag velocities are modest. A full Doppler search can be avoided by keeping the length of the code short, relative to the frequency difference of the reference clocks at the tag and receiver. These slight clock rate differences effectively create small Doppler shifts, and their impact cannot be ignored entirely. The receivers use a GPS disciplined frequency reference which offers sufficient accuracy, however the tags cannot carry a GPS receiver. Inexpensive and accurate TCXO clocks provide the tag reference frequency. Typical tolerances for these clocks are 1-5ppm, corresponding to an offset of 140 to 700Hz at 140MHz. The expected value of the base band signal as a function of carrier phase (Δθ), code phase (Δτ), and frequency offset (ΔfD) are shown in the equation below [11].
The attractive features of this system, including low power tags, automatic detection, and improved location accuracy, depend on a network of receivers that can listen continuously and in real time for tag transmissions. The real time requirement sets a hard limit on performance which impacts all other design choices. We chose to implement the matched filter detector on a Texas Instruments TMS320C6416 Digital Signal Processor. This processor can run at 1GHz and offers parallel data processing capability, with up to 8GMAC per second when operating on 8 bit data. The incoming RF signal is down-converted to base band I and Q channels and then sampled at 2.8125 MHz. An 11 bit PN sequence at the tag’s 1MHz chip rate would occupy 5760 samples, and processing the I and Q channels with a straightforward, time-domain matched filter implementation would require approximately 32 GMAC per second in order to guarantee real time operation. This requirement clearly cannot be met, even by the impressive performance of the ‘C6416, which was state of the art when this work began. The options available are to either reduce the bandwidth of the transmitted signal, which reduces the ranging accuracy, reduce the PN sequence length, which reduces the processing gain, or to use a frequency domain algorithm to implement the matched filter. Ordinary time domain correlation is an O(N2) operation, where N is the number of elements to be cross-correlated. Operation in the frequency domain in contrast is an O(Nlog10N) operation, thanks to the remarkable efficiency of the FFT. An early application of this technique to GPS was demonstrated in [12]. We chose this alternative, and used available FFT routines along with the established practice of computing a correlation by conjugate multiplication in the frequency domain, to meet the real time requirements of the system. The cross correlation function C at a particular offset in samples, k between the signals g(n) and h(n), which are both at least N samples long is
_
E{S} = C De j ( Δθ +πΔf DTCO ) R(Δτ )sinc(πΔf D TCO ) The sinc function component can be thought of as an amplitude modulating factor; larger values of ΔfD or TCO reduce the amplitude, which reduces the signal to noise ratio. While the total tag sequence is 3 msec long, the synchronization and ID codes are processed as separate halves, which yields an integration time, TCO=1.5 msec. If we set ΔfD
